JP2000305940A - Method and device for retrieving time series data and storage medium storing time series data retrieval program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform retrieval of a series of similar data with high accuracy and stability by automatically and objectively integrating dissimilarity of plural feature values capable of expressing various features of data according to an inquiry from a user and setting the total dissimilarity. SOLUTION: Time series data having a time series is inputted (S1), stored in a database, plural first featured value vectors to express a property of the time series data are calculated (S2) from the stored time series data at every interval of fixed time, an optional piece of inquiry data to be given by the user is inputted (S3) and second featured value vectors are calculated (S4) from the inquiry data. Then, the dissimilarity between the inquiry data and data of each time in the data base is calculated (S5) from the first and second featured value vectors, the dissimilarity is integrated (S6), the total dissimilarity between the inquiry data and the data of each time in the database is calculated, data of the applicable time are selected and outputted from the database.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a time-series data search method and apparatus and a storage medium storing a time-series data search program, and more particularly, to a storage medium storing a large amount of time-series data stored in a database. Application of technology to search for data similar to arbitrary time-series data provided by users, such as video database search, analysis of time-series signals that record various phenomena such as weather, earthquakes, and markets, and data mining And a storage medium storing a time-series data search program.

[0002]

2. Description of the Related Art Conventional techniques for searching for time-series data include, for example, time-series images, which are described in the documents "Yasufumi Fujimoto, Hidehiko Iwasa, Naokazu Yokoya, Haruo Takemura:" Video Search Based on Similarity ", IEICE Technical Report, PRMU 96-110, pp. 49-56, 1996.

In this method, for each frame of a time-series image,
A gray-scale value of a pixel in an image is represented as a one-dimensional feature amount vector, and a matrix is created by converting this vector calculated for each frame of the image sequence constituting the time-series image database into a column vector. A covariance matrix is obtained for a matrix obtained by subtracting the average value of the feature amount vectors, and an eigenspace based on eigenvectors corresponding to large eigenvalues obtained by solving an eigenvalue problem of this matrix is created. Calculate the trajectory projected above. The distance between the trajectory in the eigenspace corresponding to the image sequence of the search query and the trajectory in the eigenspace of the image sequence stored in the database is used as a measure of the dissimilarity between the time-series images, and the distance The image sequence having a small value is used as a search result.

When a time-series image is to be processed, in addition to the above-mentioned feature amount vector relating to the distribution of gray values of the image,
An image feature such as a motion vector field of a pattern in an image, a color, and a texture is usually used as a feature amount vector. Using these various feature vectors, the dissimilarity between the query time-series image and the time-series image in the database is calculated as the difference between the feature vectors, and the dissimilarity obtained for the various feature vectors is weighted. Conventionally, a method of searching for a time-series image having a small scale from a database by summing up the sum to obtain a total dissimilarity measure is widely used.

[0005]

However, in the above-mentioned conventional method, a plurality of feature quantity vectors such as an image density value, a motion vector field, a color, and a texture are used. When calculating the degree of dissimilarity with the data in it and weighting them to calculate the overall degree of dissimilarity, it is necessary to determine the scale of the degree of dissimilarity and determine the weighting count for each feature value There is. Traditionally, setting the dissimilarity measure is:
Mainly, the designer determines in advance empirically, and the weighting is manually set by the designer or the user.

In general, even in the same database, the characteristics to be emphasized change according to the characteristics of the data. Therefore, it is necessary to change the setting of the weighting between the feature amounts at any time during the search, and the search operation becomes difficult. It is inefficient and inefficient, and there is a problem that search accuracy is unstable and deteriorates.
Also, when narrowing down the features to a small number of features, such as the grayscale distribution of the image, the problem of determining the weight between the features can be reduced, but the ability to express various properties included in the original data is reduced, As a result, it is difficult to search for similar data desired by the user.

SUMMARY OF THE INVENTION The present invention has been made in view of the above points, and automatically determines the dissimilarity of a plurality of feature amounts capable of expressing various characteristics of data in accordance with the characteristics of inquiry data from a user. -Time data search method and apparatus and time-series data capable of performing high-accuracy, high-stability, similar data series search by integrating objectively and objectively and setting comprehensive dissimilarity It is an object to provide a storage medium storing a search program.

[0008]

FIG. 1 is a diagram for explaining the principle of the present invention. The present invention (claim 1) relates to a time-series data search method for searching for data similar to user inquiry data from a time-series database that records time-series data. A plurality of first feature vectors expressing the properties of the time-series data are calculated at regular time intervals from the stored time-series data (step 1), and stored in the database (step 2). Is input (Step 3), a second feature vector expressing the nature of the query data is calculated (Step 4), and the degree of dissimilarity between the query data and the data at each time in the database is calculated. Is calculated from the first feature amount vector and the second feature amount vector (step 5), and the dissimilarity is integrated to obtain the query data and the The overall dissimilarities between the time of the data calculated (Step 6), the increasing order of overall dissimilarity, the data of the corresponding time selected from a database, and outputs (Step 7).

According to the present invention (claim 2), when calculating the overall dissimilarity, a partial data set having a small dissimilarity is selected from the query data in the database, and the selected data set is selected. For each time in the partial data set, a first dissimilarity between the time and data at another time in the database is calculated, and a second dissimilarity between the two data after each time is calculated. , And a first dissimilarity between the data at the other times calculated for each time included in the selected partial data set, and two data after the time. Model the expected value of the second degree of dissimilarity as a function of the first degree of dissimilarity, and apply the first degree of dissimilarity to the model By integrating, query data and each time in the database To calculate the overall non-similarity between the data of.

According to the present invention (claim 3), the first dissimilarity between the inquiry data and the data at each time in the database is calculated for each feature amount vector, or two dissimilarities after the time are calculated. When calculating the second degree of dissimilarity between time data, for each of the feature amount vectors, two sequences of two corresponding time periods and two sequences of feature amount vectors corresponding to times before the current time are used. The sum or the average distance between the feature amount vectors between them is calculated while maintaining the temporal order relation.

The present invention (claim 4) provides, for each time in a selected partial data set, a first degree of dissimilarity between data after the time and data in a database. Is calculated, for any one type of feature amount vector, for two partial sequences of the feature vector corresponding to times after the two designated times, the distance between the feature amount vectors between the two sequences is calculated. The sum or the average distance is calculated while maintaining the temporal order relation.

[0012] The present invention (claim 5) provides a first dissimilarity between data at other times calculated for each time included in the selected partial data set, When modeling the relationship between the two pieces of data with the second degree of dissimilarity as a function of the first degree of dissimilarity, the expected value of the second degree of dissimilarity is modeled. The relationship between the dissimilarity between the two data and the expected value of the dissimilarity between the two data after that time is approximated using a logistic model with the first dissimilarity as a variable, and the coefficient of the model is determined by regression analysis. Ask for.

[0013] The present invention (claim 6) provides a first dissimilarity between data at other times calculated for each time included in the selected partial data set, When modeling the relationship between the two pieces of data with the second degree of dissimilarity as a function of the first degree of dissimilarity, the expected value of the second degree of dissimilarity is modeled. For a subset of data in which the second degree of dissimilarity is equal to or less than an arbitrary threshold, an expected value of the degree of dissimilarity between two pieces of data after each time is calculated, and the first threshold of dissimilarity is set as a variable Is approximated using a logistic model, and the coefficients of the model are determined by regression analysis.

According to the present invention (claim 7), a time-series image is input as time-series data. The present invention (claim 8) uses a time-series image as target data, and calculates a feature vector expressing the property of the data, and calculates a gray value of a pixel in the image for each unit time or the gray value. The average brightness value obtained by spatially averaging or the image plane is divided into meshes, and the vector is calculated as a vector having an element of the average of the grayscale values of the images included in each mesh.

According to the present invention (claim 9), when the time-series data is a time-series image, a characteristic amount vector relating to the distribution of gray values of the image is selected as one kind of characteristic amount vector. According to the present invention (claim 10), when a partial data set in the database is selected, the second feature vector of the inquiry data and the first feature vector of the data at each time in the database are selected. The distance is calculated for each of the characteristic amount vectors, and data at a time when each distance is smaller than a predetermined value is selected.

FIG. 2 is a diagram showing the principle of the present invention. The present invention (Claim 11) is a time-series data search device for searching data similar to user inquiry data from a time-series database that records time-series data. A plurality of first feature quantity vectors expressing the characteristics of the data at regular time intervals are obtained from the data series input means 20 for inputting the series data and storing the data in the database 90 and the time series data stored in the database 90. First feature value vector calculating means 20 for calculating, and search query data series input means 30 for inputting arbitrary query data given by a user
A second feature value vector calculating means 40 for calculating a second feature value vector expressing a property of the data from the inquiry data;
The first dissimilarity calculating means 50 for calculating the dissimilarity with the data at each of the time points from each of the first feature amount vector and the second feature amount vector, and the dissimilarity are integrated. Second dissimilarity calculating means 60 for calculating the overall dissimilarity between the inquiry data and the data at each time in the database.
Searching means 70 for selecting the data at the corresponding time from the database 90 in the order of the smaller overall dissimilarity.
And output means 80 for outputting the data selected by the search means 70.

According to the present invention (claim 12), in the second dissimilarity calculating means 60, means for selecting a partial data set which is a data set having a small dissimilarity from the query data in the database 90, Means for calculating, for each time in the selected data set, a first degree of dissimilarity between the time and the data at another time in the database; A first dissimilarity between the data at the other times calculated for each time included in the selected partial data set; and Means for modeling an association of the second data with the second degree of dissimilarity as a function of the first degree of dissimilarity, the expected value of the second degree of dissimilarity; By applying to the model, Have combined data and the database 90
Means for calculating the overall dissimilarity between the data at each of the times.

According to a thirteenth aspect of the present invention, in the means for calculating the first degree of dissimilarity and the means for calculating the second degree of dissimilarity, two times corresponding to each of the second feature amount vectors are provided. For two sequences of feature vectors corresponding to times before the time, a sum of distances between feature vectors between the two sequences or an average distance is calculated while maintaining a temporal order relationship. Including means.

According to the present invention (claim 14), in the means for calculating the first degree of dissimilarity, two types of feature vectors corresponding to times after two specified times are specified for any one type of feature amount vector. Means for calculating a sum of distances between feature amount vectors between two sequences or an average distance of two partial sequences while maintaining a temporal order relationship.

[0020] The present invention (claim 15) provides a first dissimilarity between data at other times calculated for each time included in the selected partial data set, Means for modeling an association of the second dissimilarity between the two data with the second dissimilarity as a function of the first dissimilarity, the expected value of the second dissimilarity, The relationship between the dissimilarity between the two data and the expected value of the dissimilarity between the two data after that time is approximated using a logistic model with the first dissimilarity as a variable, and the coefficient of the model is determined by regression analysis. Including means for determining

[0021] The present invention (claim 16) is characterized in that a first dissimilarity between data at other times calculated for each time included in the selected partial data set, Means for modeling an association of the second dissimilarity between the two data with the second dissimilarity as a function of the first dissimilarity, the expected value of the second dissimilarity, Second in between
For a subset of data in which the degree of dissimilarity is equal to or less than an arbitrary threshold, an expected value of dissimilarity between two pieces of data after each time is calculated, and a logistic model using the first threshold of dissimilarity as a variable And means for obtaining the coefficients of the model by regression analysis.

According to the present invention (claim 17), the data series input means 10 inputs a time series image as time series data. According to the present invention (claim 18), when a time-series image is input as time-series data in the first feature amount vector calculation means 20, a gray scale value of a pixel in the image per unit time, or Means for calculating an average luminance value obtained by spatially averaging the grayscale values or a vector having an image plane divided into meshes and having an average of grayscale values of images included in each mesh as an element.

According to a nineteenth aspect of the present invention, in the means for calculating the first degree of dissimilarity, when the time-series image is time-series data, the gray-scale value Means for selecting a feature quantity vector related to the distribution of. According to the present invention (claim 20), in the means for selecting a partial data set in the database, the second feature vector of the inquiry data and the first feature vector of the data at each time in the database are used. The method includes means for calculating the distance for each of the characteristic amount vectors and selecting data at a time when each distance is smaller than a predetermined value.

The present invention (claim 21) is a storage medium storing a time-series data search program for searching data similar to user inquiry data from a time-series database for recording time-series data. A data series input process for inputting time series data forming a time series and configuring the data series, and a plurality of first series expressing the properties of the data at regular time intervals based on the time series data stored in the database. A first feature amount vector calculation process for calculating a feature amount vector, a search query data series input process for inputting arbitrary query data given by a user, and a second process for expressing a property of the data from the query data A second feature vector calculation process for calculating the feature vector; The first dissimilarity calculation process of calculating the dissimilarity with the data at each time in the database from each of the first feature amount vector and the second feature amount vector, and the dissimilarity are integrated. A second dissimilarity calculation process for calculating the overall dissimilarity between the inquiry data and the data at each time in the database; and It has a search process to select more and an output process to output the data selected by the search process.

According to the present invention (claim 22), in the second dissimilarity calculation process, a process of selecting a partial data set which is a data set having a small dissimilarity from the query data in the database is selected. Calculating, for each time in the collected data set, a first dissimilarity between the time and the data at another time in the database; and a second dissimilarity between the two data after the respective time. A process for calculating the dissimilarity, a first dissimilarity between the data of the other times calculated for each time included in the selected partial data set, and two Modeling the association between the data with the second dissimilarity as a function of the first dissimilarity, and modeling the expected value of the second dissimilarity; To apply to Ri integrated, and a process of calculating the overall dissimilarities between data at each time in the inquiry data and the database.

According to the present invention (claim 23), in the process of calculating the first dissimilarity and the process of calculating the second dissimilarity, two times corresponding to each of the second feature amount vectors and For two sequences of feature vectors corresponding to times before the time, a sum of distances between feature vectors between the two sequences or an average distance is calculated while maintaining a temporal order relationship. Including processes.

According to the present invention (claim 24), in the process of calculating the first degree of dissimilarity, for any one type of feature amount vector, two of feature vectors corresponding to times after two specified times are specified. The method includes a process of calculating a sum of distances between feature value vectors between two sequences or an average distance of two partial sequences while maintaining a temporal order relationship.

[0028] The present invention (claim 25) provides a first dissimilarity between data of other times calculated for each time included in the selected partial data set, In a process of modeling the expected value of the second dissimilarity as a function of the first dissimilarity as a function of the first dissimilarity between the two data of The relationship between the dissimilarity between the two data and the expected value of the dissimilarity between the two data after that time is approximated using a logistic model with the first dissimilarity as a variable, and the coefficient of the model is determined by regression analysis. Including the process of seeking.

The present invention (claim 26) provides a first dissimilarity between data of another time calculated for each time included in the selected partial data set, In a process of modeling the expected value of the second dissimilarity as a function of the first dissimilarity as a function of the first dissimilarity between the two data of For a subset of data in which the second degree of dissimilarity is equal to or less than an arbitrary threshold, an expected value of the degree of dissimilarity between two pieces of data after each time is calculated, and the first threshold of dissimilarity is set as a variable Approximation using a logistic model, and obtaining the coefficients of the model by regression analysis.

According to the present invention (claim 27), in the data series input process, a time series image is inputted as time series data. According to the present invention (claim 28), in the first feature amount vector calculation process, when a time-series image is input as time-series data, the gray value of a pixel in the image or the gray value of the pixel in each unit time is obtained. The method includes a process in which an average brightness value obtained by spatially averaging the values or an image plane is divided into meshes, and the vector is calculated as an element having an average of the grayscale values of the images included in each mesh.

According to the present invention (claim 29), in the process of calculating the first degree of dissimilarity, when the time-series image is time-series data, the gray-scale value And selecting a feature vector related to the distribution of. According to the present invention (claim 30), in the process of selecting a partial data set in the database, the distance between the feature vector of the query data and the feature vector of the data at each time in the database is determined.
The method includes a process of calculating each of various feature amount vectors and selecting data at a time at which each distance is smaller than a predetermined value.

As described above, in the present invention, the data at each time of the time-series data is converted into a plurality of feature vectors, and the degree of similarity between the original data is determined based on the difference between the feature vectors. . Therefore, when an enormous amount of data such as a video is to be processed, processing is performed using a feature amount vector having a significantly smaller number of dimensions as the dimensions of the original data, so that the calculation cost required for the search can be reduced. In addition, by selecting in advance the desired features representing various characteristics of the data in accordance with the type of data and the intended use, it is possible to compare the similarity between the original data having a large degree of redundancy and the similarity directly. Therefore, it is possible to perform an accurate search that meets the intended use, and by presenting the degree of similarity of each feature to the user at the time of the search, the user can grasp each characteristic of the searched data. It becomes possible.

Further, according to the present invention, the dissimilarity between the query data and the data in the database is calculated for a plurality of feature quantity vectors, and the dissimilarities are integrated to form a comprehensive data of the entire data. Dissimilarity is calculated.
Therefore, the user does not need to manually set the weighting of the dissimilarity of various feature amounts, and the search can be performed efficiently and at high speed. In addition, there is no difference in search results due to individual differences among users, and a search can always be executed with high accuracy.

Further, according to the present invention, when weighting of the degree of dissimilarity of various feature amounts is set, the dissimilarity of each feature amount between two data at the search target time and the dissimilarity degree at subsequent times are set. It models the relationship between the expected value of
Therefore, it is possible to preferentially search for the time-series data in the database that also matches the time change after the time of the inquiry data.

Further, in the present invention, when modeling the above relationship, by selecting a set of data series in the database similar to the inquiry data, the change tendency is similar to the time series data of the inquiry. Since modeling can be performed using the data series in the database, it is possible to use a search scale that matches the characteristics of the inquiry data at each time.

Further, according to the present invention, one feature amount vector which is most important for the purpose of search is selected, and the future value of the selected feature amount vector for the search query time series data and the time series data of the search candidate are selected. The search candidates can be selected in the order in which the average dissimilarity with the future value of the feature amount vector is expected to be small. For this reason, we search past similar time series data stored in the database from the time series data observed at the present time, and refer to the time change of the searched time series data to determine future changes in the current data. In the usage application of prediction, by selecting a feature amount representing the property that is most important as the feature amount vector, it becomes possible to search the database for time-series data with similar future values for the feature amount.

For example, when a time series of a weather data image obtained by observing a precipitation phenomenon is to be processed, a spatial distribution of shading of a weather radar image, a velocity vector field, a feature vector such as a texture of a pattern surface, and the like are used. By selecting a feature vector related to the spatial distribution of shades as the above-described feature vector and performing a search, a past radar image having a similar spatial distribution of shades in the future can be searched. It is possible to anticipate changes in weather.

Further, in the present invention, the relation between the dissimilarity between data at two times and the dissimilarity between two data after that time is determined by using a logistic model by using an expected value of the latter as a function of the former. Modeled. Therefore, as the dissimilarity of each feature increases, the property that the dissimilarity between the data after the time becomes larger,
It can be approximated with high accuracy. In addition, using a dissimilarity of each feature amount vector or a logistic model having a threshold value imposed on the dissimilarity as a variable, the count of each variable is determined by a regression distribution. Therefore, by performing a test of significance for each count, it is possible to select a useful feature amount vector.

[0039]

FIG. 3 shows a configuration of a time-series data search device according to the present invention. The time-series data search device shown in FIG. 1 includes an input unit 100, a processing unit 200, and an output unit 300. The input unit 100 includes a data sequence input unit 101 for inputting a time-series data sequence to be stored in a database, and a search query data sequence input unit 102 for inputting a time-series data sequence for a search query.

The processing section 200 includes a data series input section 101
, A feature extraction unit 202 that calculates a feature vector of data from the data sequence stored in the original data sequence storage unit 201 at regular time intervals, and a feature extraction unit A feature amount series storage unit 203 that stores various feature amount vectors of the data calculated in 202, a search scale setting that sets a similarity scale between query data and data in a database (original data sequence storage unit 201). The unit 204 receives a feature amount vector input from the search query data sequence input unit 102 and calculated by the feature extraction unit 202, and converts a data sequence stored in the original data sequence storage unit 201 similar to the query data into a feature. Search that performs a search based on the criteria set by the search scale setting unit 204 as a function of the dissimilarity of the quantity vector Consisting of 205.

The output unit 300 outputs the image sequence of the search result output from the processing unit 200 to a display device or a file device. FIG. 4 is a flowchart showing an outline operation of the time-series data search device of the present invention. Step 101) First, a time series data set (query data) for database construction is input from the data series input unit 101, and the original data series storage unit 20 of the processing unit 200 is input.
Accumulate in 1.

Step 102) The feature vector is calculated at regular time intervals from the time-series data input in the feature extracting unit 201 and stored in the feature-series storing unit 203.
Step 103) Search query data series input unit 1
Enter time series data for search inquiry from 02,
The feature amount vector is calculated in the feature extracting unit 201.

Step 104) Search scale setting section 204
In the query data input from the search query data sequence input unit 101 and the original data sequence storage unit 201
A search similarity measure is set from the data series. Step 105) The search unit 205 searches the database for data similar to the inquiry data series, and the output unit 3
Output from 00.

[0044]

Embodiments of the present invention will be described below with reference to the drawings. In the following, as an example of the time-series data, the operation of the time-series data search device for a time-series image will be described.
A specific description will be made focusing on 0. The original data sequence storage unit 201 in FIG. 3 is a database that stores time-series data sequences input through the data sequence input unit 101. Here, a time-series image is targeted, and a set of a series of image series that are continuous in time and an image series that is continuous in an arbitrary section can be input and stored. By specifying the frame time or the frame number, the corresponding image can be taken out.

The feature extraction unit 202 expresses the nature of data at regular time intervals from the data sequence stored in the original data sequence storage unit 201 and the data sequence input through the search query data sequence input unit 102. Calculate a plurality of feature vectors. Here, as an example of a case of a time-series image, a mesh feature representing a spatial distribution of gray values of a pattern in an image, a velocity vector field representing a motion distribution of a pattern, and fine movements and textures of a pattern surface are described. Temporal texture feature 3 to express
An example of calculating two feature vectors will be described.

FIG. 5 is a diagram for explaining an example of a feature vector when a time-series image is used as target data according to an embodiment of the present invention. The mesh feature is obtained at a certain time t by using an image I (i, j, t) obtained as shown in FIG.
Is a one-dimensional feature amount vector x ₁ (t) that is divided into meshes as shown in FIG. 5B, and that averages the gray values i (i, j, t) of the pixels in each mesh. FIG.
An average velocity vector of a pattern included in each mesh as shown in FIG. 5B is calculated as a velocity vector field as shown in FIG. 5C, and a one-dimensional vector having each component as an element is calculated as a velocity vector field. Let it be a feature vector x ₂ (t).

Further, the temporal texture feature is an image feature that expresses a fine movement or texture of a pattern expression, and the nature of a non-rigid movement pattern accompanied by generation and disappearance of a pattern can be quantified.
It calculates the probability density distribution of the motion components of the patterns contained in multiple image frames, and from that distribution, the variety of motions contained in the local spatio-temporal region of the time-series image and the rules for the arrangement of image elements Image features such as gender are calculated.
Here, a vector having the elements such as the magnitude of the predominant speed, the uniformity of the movement, the concealment ratio, the direction of the contour arrangement, the roughness of the contour arrangement, the contrast of the contour, etc. is calculated at regular time intervals, The feature amount vector is x ₃ (t).

The feature vector obtained for the data contained in the original data series storage unit 201 is stored in the feature series storage unit 203. At this time, the variance and the average of each feature amount vector are normalized, and the same parameter is used for the feature amount vector obtained from the query data sequence input from the search query data sequence input unit 101. Perform normalization with.

The feature amount series storage unit 203 stores the feature amount vector calculated by the feature extraction unit 202 and, in response to a request from the search scale setting unit 204 and the search unit 205, a feature corresponding to the requested time. Output a quantity vector. The search scale setting unit 204 calculates the degree of dissimilarity of the query data calculated by the feature extraction unit 203 with the feature vector, for each of the various feature vectors.
By integrating the dissimilarities, a search scale is set to calculate the overall dissimilarity between the inquiry data and the data at each time in the database.

Here, the following method will be described as an example. Now, as the inquiry data, the time T
, T−1, T} at L time steps including the observed image. Here, this data series is denoted by T. Therefore, the data at time T and the original data series storage unit 2
01, the partial data sequence {t−L + 1,..., T−1, t} at time T, and the similarity D (T, t) = {D _k (T , T) | k = 1, 2, 3}. In this example, a method using DP matching is used to correct the time expansion and contraction between two partial sequences, and the calculation is performed as follows.

[0051]

(Equation 1)

Where η is a constant and | · |
Donorm. V_k(T) is the query data series
Is the speed of the pattern change._k(T) = | x_k(T) -x_k(T-1) | (2) This V_kUsing (T), two
Normalized non-independent
Similarity d_kCalculate (m, n) and query it
For the step width L of the data time
Like dissimilarity D _kCalculate (T, t). This dissimilarity
The degree calculation method is also used in the search unit 205.

Next, using the method defined above, the degree of dissimilarity between the query T and the data at each time in the original data series storage unit 201 is determined by using the respective feature amount vectors k
Is calculated. Then, as the candidate data sequences to be searched, the non-similarity of each feature vector, the threshold xi] = user gives _{{ξ 1, ξ 2, ξ} 3} finding a set S _T of the data sequence satisfying.

[0054]

(Equation 2)

Next, with reference to each element in the search candidate set S _T, by integrating the non-similarity of each feature quantity vectors, sets a measure for determining the overall dissimilarities. An example of the method will be described below. First, each time t in the search candidate set S _T
For ∈S _T, calculate the dissimilarity D _k between the data with other time tau in the original data sequence storage unit 201 (t, tau), also the same time, that time t, tau subsequent The dissimilarity y (t, τ) between the two sequences at the time is calculated for the feature amount vector of interest. This dissimilarity y (t,
τ) is called a prediction error. Here, the above-described feature extraction unit 2
The mesh feature x ₁ (t) described as an example of the feature vector of 02 is set as a feature vector of interest. Therefore, the prediction error y (t, τ) is

[0056]

(Equation 3)

The calculation is as follows. Next, from the set of the dissimilarity D _k (t, τ) obtained for each element in the data set Sτ thus selected and the prediction error y (t, τ), the dissimilarity is imposed. Threshold value ε = (ε ₁ , ε ₂ , ε ₃ )
Expected value e of the prediction error y (t, τ) for the set satisfying
Calculate (ε).

[0058]

(Equation 4)

Further, the expected value e (ε) of the prediction error
Is regarded as a function of the threshold ε of dissimilarity, and modeling is performed using regression analysis. Here, an example in which a logistic model is used as one of the models will be described.

[0060]

(Equation 5)

Here, g (ε) is a polynomial using the threshold value ε _k as a variable, and b ₀ is one coefficient. The expected value of the prediction error calculated using such a model

[0062]

(Equation 6)

Is referred to as an expected prediction error. The expected prediction error refers to an estimated value of an average dissimilarity regarding a feature amount focused on future data and a future value of inquiry data for a set of search candidates satisfying a certain threshold ε. When all the properties of the data at each time in the database are uniform, the search candidate set Sτ used when modeling the expected value of the prediction error can be calculated as all the data in the database.

The model of the equation (6) thus obtained is output to the search unit 205. The search unit 205 selects a data sequence in the original data sequence storage unit 201 that is similar to the query data sequence input from the search query data sequence input unit 102, and outputs the data sequence to the output unit 300. Here, the following method is shown as an example.

As described above in the search scale setting unit 204,
Now, from the search candidates in the search candidate set Sτ obtained for the query data sequence at time T, the query data sequence is set using the search scale (expected prediction error model) set by the search scale setting unit 204 described above. And the expected prediction error between the search candidate and the data series of the search candidate is calculated as a total degree of dissimilarity, the search candidates are sorted in a similar order, a search candidate that satisfies the conditions specified by the user is selected, and the search is performed. The result is output to the output unit 300.

Here, an example of the condition specified by the user is shown.
And the maximum number K of data series of the search result
The expected prediction error threshold that the set of data series should satisfy
Value e _THAnd find the data series of the search results as follows
You.

[0067]

(Equation 7)

A specific example for a weather radar image will be described below. FIG. 6 shows an example of a search query data sequence provided by a user according to an embodiment of the present invention. The figure shows the beginning and end frames of a time-series image input as a search query data sequence given by a user. However, the unit time step is one hour, and the search inquiry data sequence includes data for three hours and is composed of 36 frames.

As described in the description of the feature extracting unit 202, mesh features, velocity vector fields,
Three types of temporary texture features were used. Also, the original data series storage unit 201 stores about 9000 hours (1
(20,000 frames). The search scale set by the search scale setting unit 204 for the inquiry data of FIG.

[0070]

(Equation 8)

FIG. 7 shows this state. The points in the figure are the calculated sample points, and the correlation coefficient is 0.989 as compared with the model represented as a curved surface, indicating that the modeling has been performed well. FIG. 8 shows the search results obtained for the search query data series in FIG. The time-series image, which is search query data input by the user, shows a linear pattern approaching the land while slightly distorting its shape. It can be confirmed that a pattern having a similar tendency was obtained.

The present invention includes a hard disk capable of storing data and freely reading the data, a device similar thereto, and a device such as a display device for displaying and outputting desired information. Based on a computer having a central processing unit or the like that controls based on the procedure or a device equivalent thereto, a part or all of the processing of each unit in the above-described embodiment, or the procedure shown in the flowchart in FIG. The present invention can be realized by giving a processing program describing an algorithm or a program similar thereto to the computer, and controlling and executing the processing program. Here, a CD-ROM, floppy disk (FD),
It can be recorded on a magneto-optical disk (MO) or a storage medium equivalent thereto and distributed.

The present invention is not limited to the above embodiment, but can be variously modified and applied within the scope of the claims.

[0074]

As described above, according to the present invention, in searching for similar time-series data, a feature vector expressing the property of the data is calculated, and the query data and the data in the database are dissimilar. The degree is calculated for each feature vector, they are automatically and objectively integrated, and the overall dissimilarity between the query data and the data in the database is set. By selecting similar time series data from
Simple and efficient search with high accuracy is possible even for complex data such as a video of a natural phenomenon.

Further, in the above method of integrating the dissimilarity of each feature amount vector, one of the plurality of feature amount vectors is selected as the most important feature amount vector,
Since the future dissimilarity between the query data and the search candidate data is modeled as a function of the dissimilarity for various feature vectors, similar retrieval is possible not only for the current time series data but also at a future time. It is possible to select a candidate, which is effective for use in predicting a change in future data based on past data.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a principle configuration diagram of the present invention.

FIG. 3 is a configuration diagram of a time-series data search device of the present invention.

FIG. 4 is a flowchart showing an outline operation of the time-series data search device of the present invention.

FIG. 5 is a diagram for describing an example of a feature amount vector when a time-series image is set as target data according to an embodiment of the present invention.

FIG. 6 is an example of a search query data sequence provided by a user according to an embodiment of the present invention.

FIG. 7 is a diagram for explaining a search scale according to an embodiment of the present invention.

FIG. 8 is an example of a data sequence of a search result obtained for an inquiry data sequence according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 Data series input means 20 First feature vector calculation means 30 Search query data series input means 40 Second feature vector calculation means 50 First dissimilarity calculation means 60 Second dissimilarity calculation means 70 Search Means 80 Output means 90 Database 100 Input unit 101 Data sequence input unit 102 Search query data sequence input unit 200 Processing unit 201 Original data sequence storage unit 202 Feature extraction unit 203 Feature amount sequence storage unit 204 Search scale setting unit 205 Search unit 300 Output Department

 ────────────────────────────────────────────────── ─── Continuing from the front page (72) Inventor Haruhiko Kojima 3-19-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo F-term in Nippon Telegraph and Telephone Corporation (reference) 5B075 ND12 NK07 PQ02 PR06 UU40 5L096 FA32 FA37 FA66 FA67 GA19 GA51 HA04 HA08 JA03 JA11

Claims (30)

[Claims] 1. A time series data search method for searching data similar to user inquiry data from a time series database that records time series data, wherein time series data forming a time series is input and stored in the database. Calculating a plurality of first feature vectors expressing the properties of the time-series data at predetermined time intervals from the stored time-series data, inputting arbitrary inquiry data given by a user, Calculating a second feature vector from the query data; calculating a dissimilarity between the query data and the data at each time in the database from the first feature vector and the second feature vector; Integrating the dissimilarities to calculate a total dissimilarity between the query data and the data at each time in the database; In increasing order of overall dissimilarity, the data of the corresponding time selected from the database, sequence data search method when and outputting. 2. When calculating the overall dissimilarity, a partial data set having a small dissimilarity is selected from query data in the database, and the selected partial data is selected. For each time in the set, calculate a first dissimilarity between the time and the data at another time in the database; a second dissimilarity between the two data after each time And the first dissimilarity between the data at the other times calculated for each time included in the selected partial data set, and the two data after the time. Modeling an association between the second dissimilarity and the expected value of the second dissimilarity as a function of the first dissimilarity, and applying the first dissimilarity to the model The inquiry data and the data 2. The time-series data search method according to claim 1, wherein the overall dissimilarity between data at each time in the database is calculated. 3. When calculating the first degree of dissimilarity between the inquiry data and the data at each time in the database for each feature amount vector, or between data at two times after the time. When calculating the second degree of dissimilarity of each of the feature amount vectors,
For two sequences of feature vectors corresponding to one time and a time before the time, the sum of the distances between the feature vectors between the two sequences or the average distance is stored in a temporal order relationship. 3. The time-series data search method according to claim 1, wherein the time-series data is calculated. 4. When calculating the first dissimilarity between data after the time and data in the database for each time in the selected partial data set. In addition, for any one type of feature amount vector, 2
For two subsequences of the feature vector corresponding to times after two times, the sum of the distances between the feature vectors between the two sequences or the average distance is calculated while maintaining the temporal order relation. The time-series data search method according to claim 2. 5. The first degree of dissimilarity between data at other times calculated for each time included in the selected partial data set, and two data after the time. When modeling the association of the second dissimilarity with the second dissimilarity as a function of the first dissimilarity, the expected value of the second dissimilarity is modeled. The relationship between the degree and the expected value of the degree of dissimilarity between the two pieces of data after the time is represented by the first
3. An approximation is made using a logistic model having a degree of dissimilarity as a variable, and coefficients of the model are obtained by regression analysis.
The described time-series data search method. 6. The first degree of dissimilarity between data at other times calculated for each time included in the selected partial data set, and two data after the time. Modeling the association of the second dissimilarity with the second dissimilarity as a function of the first dissimilarity when the expected value of the second dissimilarity is modeled. For a subset of data in which the degree of dissimilarity of 2 is equal to or less than an arbitrary threshold, an expected value of the degree of dissimilarity between two pieces of data after each time is calculated.
3. The time-series data search method according to claim 2, wherein approximation is performed using a logistic model having a threshold value of the degree of dissimilarity as a variable, and coefficients of the model are obtained by regression analysis. 7. The time-series data search method according to claim 1, wherein a time-series image is input as the time-series data. 8. When the time-series image is used as target data to calculate a feature vector expressing the property of the data, the grayscale value of a pixel in the image or the grayscale value is spatially determined for each unit time. 8. The time-series data search method according to claim 7, wherein the averaged average luminance value or the image plane is divided into meshes, and the vector is calculated as a vector having an average of the gray values of the images included in each mesh as elements. 9. The image processing apparatus according to claim 4, wherein, when the time-series data is a time-series image, a characteristic amount vector relating to a distribution of gray values of the image is selected as the arbitrary one type of characteristic amount vector.
The described time-series data search method. 10. When selecting a partial data set in the database, the second feature vector of the inquiry data;
3. The distance between the data at each time in the database and the first feature value vector is calculated for each of the various feature value vectors, and data at a time at which each distance is smaller than a predetermined value is selected. Time series data search method. 11. A time-series data search device for searching data similar to user inquiry data from a time-series database that records time-series data, wherein time-series data forming a time series constituting the database is input. A data sequence input unit to be stored in the database; and a first feature for calculating a plurality of first feature amount vectors expressing the properties of the data at regular time intervals from the time series data stored in the database. Quantity vector calculation means, search query data series input means for inputting arbitrary query data given by the user, and a second feature quantity vector expressing a property of the data from the query data. A feature vector calculation means, and a degree of dissimilarity between the inquiry data and data at each time in the database. A first dissimilarity calculating means for calculating from the first feature amount vector and the second feature amount vector, and integrating the dissimilarity to obtain the query data and the time of each time in the database. Second dissimilarity calculating means for calculating overall dissimilarity with data; search means for selecting data at the corresponding time from the database in the order of the small overall dissimilarity; Output means for outputting data selected by the search means. 12. The second dissimilarity calculating means: means for selecting a partial data set that is a data set having a small dissimilarity from the query data in the database; Means for calculating a first dissimilarity between said time and data at another time in said database; a second dissimilarity between two data after each said time Means for calculating a degree, a first dissimilarity between data of other times calculated for each time included in the selected partial data set, and 2 after the time. Means for modeling an association of the second data with the second degree of dissimilarity as a function of the first degree of dissimilarity, the expected value of the second degree of dissimilarity; By applying to the model , The time-series data retrieval apparatus according to claim 11, further comprising a means for calculating the overall dissimilarities between data at each time in the said inquiry data database. 13. The means for calculating the first degree of dissimilarity and the means for calculating the second degree of dissimilarity include two corresponding times for each of the second feature amount vectors and a time before the time. Means for calculating the sum of the distances between the feature vectors between the two sequences or the average distance of the two sequences of the feature vectors corresponding to the time while maintaining the temporal order relationship. Item 13. The time-series data search device according to item 11 or 12. 14. The means for calculating the first degree of dissimilarity may include:
For two subsequences of the feature vector corresponding to times after two times, the sum of the distances between the feature vectors between the two sequences or the average distance is calculated while maintaining the temporal order relation. 13. The time-series data search device according to claim 12, comprising means. 15. The first degree of dissimilarity between data at other times calculated for each time included in the selected partial data set, and two data after the time. Means for modeling the expected value of the second dissimilarity as a function of the first dissimilarity as a function of the first dissimilarity. The relationship between the degree and the expected value of the degree of dissimilarity between the two pieces of data after the time is represented by the first
13. The time-series data search device according to claim 12, further comprising means for approximating using a logistic model that uses the degree of dissimilarity as a variable, and obtaining a coefficient of the model by regression analysis. 16. The first degree of dissimilarity between data at other times calculated for each time included in the selected partial data set, and a difference between two data after the time. Means for modeling the association of the second dissimilarity with the second dissimilarity as a function of the first dissimilarity, wherein the expected value of the second dissimilarity is: For a subset of data in which the degree of dissimilarity of 2 is equal to or less than an arbitrary threshold, an expected value of the degree of dissimilarity between two pieces of data after each time is calculated.
13. The time-series data search device according to claim 12, further comprising means for approximating using a logistic model having a threshold value of the degree of dissimilarity as a variable and obtaining a coefficient of the model by regression analysis. 17. The data series input unit inputs a time series image as the time series data.
The time-series data search device according to claim 2, 13, or 14. 18. When a time-series image is input as the time-series data, the first feature amount vector calculation means calculates a gray-scale value of a pixel in the image for each unit time or a gray-scale value of the pixel. 12. Time-series data according to claim 11, further comprising means for calculating an average brightness value obtained by spatially averaging or dividing the image plane into meshes and calculating as a vector having an element of an average of the gray values of the images included in each mesh as an element. Search device. 19. The means for calculating the first degree of dissimilarity, wherein when the time-series image is the time-series data, the one or more types of feature amount vectors are characterized by a distribution of gray values of the image. 2. The method of claim 1, further comprising means for selecting a quantity vector.
4. The time-series data search device according to 4. 20. A means for selecting a partial data set in the database, the second feature amount vector of the inquiry data;
Means for calculating a distance between the data at each time in the database and the first feature value vector for each of the various feature value vectors, and selecting data at a time at which each distance is smaller than a predetermined value. Item 13. The time-series data search device according to Item 12. 21. A storage medium storing a time-series data search program for searching for data similar to user inquiry data from a time-series database that records time-series data, the time-series data constituting a database. A data sequence input process for inputting time-series data and storing the data in a database; and calculating a plurality of first feature vectors expressing the properties of the data at regular time intervals from the time-series data stored in the database. A first feature amount vector calculation process for inputting, a search query data series input process for inputting arbitrary query data given by the user, and a second feature amount vector expressing a property of the data from the query data. A second feature amount vector calculation process to be calculated; A first dissimilarity calculation process of calculating a dissimilarity with data at each time in the database from each of the first feature amount vector and the second feature amount vector; and integrating the dissimilarity. A second dissimilarity calculation process for calculating the overall dissimilarity between the inquiry data and the data at each time in the database; A storage medium storing a time-series data search program, comprising: a search process for selecting the data from the database; and an output process for outputting the data selected by the search process. 22. The second dissimilarity calculation process includes: selecting a partial data set that is a data set having a low dissimilarity from the query data in the database; Calculating a first degree of dissimilarity between the time and data at other times in the database for each time of the second time; and a second dissimilarity between two data after the respective time. Calculating the degree of dissimilarity; the first degree of dissimilarity between the data of the other times calculated for each time included in the selected partial data set; Modeling an association of a second dissimilarity between the two data as a function of the first dissimilarity with the expected value of the second dissimilarity; To the model Time-series data retrieval program storage medium storing according to claim 21, wherein said having a process of calculating the overall dissimilarities between data at each time in the integrated, the and the inquiry data database by. 23. The process of calculating the first dissimilarity and the process of calculating the second dissimilarity include two times corresponding to each of the second feature amount vectors and a time before the time. Claims including a process of calculating a sum of distances between feature vectors between two sequences or an average distance of two sequences of feature vectors corresponding to time while maintaining a temporal order relationship. Item 21
Alternatively, a storage medium storing the time-series data search program according to 22. 24. The process of calculating the first degree of dissimilarity comprises the steps of:
For two subsequences of the feature vector corresponding to times after two times, the sum of the distances between the feature vectors between the two sequences or the average distance is calculated while maintaining the temporal order relation. 23. A storage medium storing a time-series data search program according to claim 22, including a process. 25. The first degree of dissimilarity between data at other times calculated for each time included in the selected partial data set and two data after the time. Modeling the association of the second dissimilarity with the second dissimilarity as a function of the first dissimilarity, the dissimilarity between the data at two times The relationship between the degree and the expected value of the degree of dissimilarity between the two pieces of data after the time is represented by the first
23. The storage medium storing the time-series data search program according to claim 22, including a process of approximating using a logistic model having a degree of dissimilarity as a variable and obtaining a coefficient of the model by regression analysis. 26. The first degree of dissimilarity between data at other times calculated for each time included in the selected partial data set, and the two data after the time. Modeling the association of the second dissimilarity with the second dissimilarity as a function of the first dissimilarity, wherein the expected value of the second dissimilarity is: For a subset of data in which the degree of dissimilarity of 2 is equal to or less than an arbitrary threshold, an expected value of the degree of dissimilarity between two pieces of data after each time is calculated.
23. The storage medium storing the time-series data search program according to claim 22, comprising a process of approximating using a logistic model having a threshold value of the dissimilarity as a variable and obtaining a coefficient of the model by regression analysis. 27. The data series input process, wherein a time series image is inputted as the time series data.
24. A storage medium storing the time-series data search program according to claim 1, 22, or 23. 28. The first feature amount vector calculation process, wherein, when a time-series image is input as the time-series data, a gray-scale value of a pixel in the image for each unit time, or the gray-scale value 22. The time-series data according to claim 21, further comprising a process of dividing an image plane into an average brightness value obtained by spatially averaging or dividing the image plane into a mesh and calculating a vector having an element of an average of gray values of an image included in each mesh as an element. A storage medium storing a search program. 29. The process of calculating the first degree of dissimilarity, comprising: when the time-series image is the time-series data, as the arbitrary one type of feature amount vector; 25. The storage medium storing the time-series data search program according to claim 24, comprising a process of selecting a feature vector related to the time series data. 30. A process for selecting a partial data set in the database, comprising: calculating a distance between a feature vector of the inquiry data and a feature vector of data at each time in the database; 23. The storage medium storing the time-series data search program according to claim 22, including a process of calculating each of the time series and selecting data at a time at which each distance is smaller than a predetermined value.